What Happens When the AI Boom Runs Out of Money
In a Nutshell
The AI boom risks repeating the railroad era's capital destruction: $800B-1.3T in compute spending faces payback mismatches where revenue may not materialize before capital markets lose patience, even as long-term value persists. TSMC's risk transfer to hyperscalers creates concentrated single-point failure, while commodity dynamics in chips and power will eventually compress Nvidia's margins as Amazon/Google custom chips and abundant energy erode differentiation. The sustainable winners will be those with zero-marginal-cost distribution (Google/Meta ads) or internal demand to iterate custom silicon (Amazon), while pure frontier labs face the classic subscription-to-advertising pivot when consumer monetization fails.
These notes were generated by AI and may contain inaccuracies.
The speaker argues that a US victory in the AI race would be problematic from a game theory perspective. In a scenario where the US achieves meaningful military superiority through AI, China's optimal response would be to destroy TSMC facilities. The concern stems from a fundamental disconnect between Silicon Valley rhetoric and the realities of national security implications.
The speaker identifies significant dependency on China for manufacturing components like fabs and actuators, which cannot be resolved without conflict. Economic incentives prevent companies from moving production to the US because competitors sourcing from China would maintain cost advantages. Apple provides an example of limited diversification to India while remaining substantially dependent on China.
The current equilibrium appears favorable to the US, with OpenAI and Anthropic on the frontier, Google maintaining relevance, and Chinese models distilling capabilities to remain 6-9 months behind. The speaker questions whether this balance can be sustained, particularly as AI improves itself and potentially accelerates development.
Open source models are not truly free because inference costs remain substantial. GLM and Kimi incur significantly higher per-answer costs compared to expectations. The distinction between R&D costs and inference costs is critical to understanding the economics of AI deployment.
The speaker expresses concern that safety concerns may lead to reduced public releases, creating information asymmetry about frontier capabilities. The timing mismatch between return on investment and capital requirements presents a significant risk, with companies progressing from free cash flow through debt markets to equity issuance.
The railroad era demonstrated how infrastructure investments with long duration mismatches could exhaust available capital. Despite the 1870s financial crisis, railroads continued operating and contributing to GDP, with railroad money from Berkshire Hathaway now flowing into Google through the Nvidia deal.
Berkshire Hathaway's acquisition of BNSF railway illustrates the transition from high-margin, limited-scale businesses to lower-margin, high-absolute-profit opportunities. The See's Candies profits were redirected into a business with larger absolute dollar amounts despite lower margins.
Google's search business represents a high-margin, zero-marginal-cost aggregator model. The AI opportunity, while cash-intensive, offers potentially larger absolute profits. The speaker suggests Berkshire Hathaway's investment may serve as both financial backing and strategic model for Google's evolution.
The speaker distinguishes between verifiable domains like coding and mathematics versus unverifiable domains with long verification loops. The labs' responses comparing AI to solving chess or Go are critiqued as these represent knowable, bounded domains rather than genuinely new capabilities.
AI models trained on internet data capture the end state of human thought rather than the underlying thought processes. Neuralink is discussed as potentially capturing thought traces that could expand model capabilities beyond current limitations.
Even without further improvements, AI's current capabilities represent massive economic opportunity across numerous jobs and economic activities. The speaker identifies as a reluctant accelerationist, arguing for continued development while acknowledging timing uncertainties.
Medicine represents one of the largest AI opportunities due to the potential for machine learning analysis of medical records to generate discoveries and improved treatments. However, regulatory barriers and access issues present significant challenges to implementation.
Aggregation theory emphasizes zero marginal cost advantages, particularly in distribution where the challenge is demand discovery rather than distribution itself. AI changes this dynamic significantly because inference costs are real and variable depending on usage patterns.
The marginal cost of serving different AI users varies dramatically, from simple queries comparable to serving web pages to complex problem-solving that may require days or weeks of computation. Microsoft's enterprise pricing model reflects this challenge with usage-based charges beyond base licensing fees.
Traditional enterprise budgeting operates on annual cycles and headcount-based licensing. Usage-based AI pricing introduces monthly spending decisions that most companies are not structured to handle, creating friction in adoption and revenue predictability for providers.
Microsoft faces a fundamental tension in pricing AI services for enterprises. Traditional capex thinking leads companies to make one-time decisions about employee costs and then move on, but AI usage requires ongoing monthly evaluation of actual consumption. This creates pressure for customers to scrutinize their Microsoft bills and question the value of each product. The extreme users who consume massive token volumes cost Microsoft far more than the flat $100 monthly rate, making them unsustainable to support, yet Microsoft needs to retain the fixed pricing model for the majority of employees who fit within standard usage patterns.
The recipe builder user represents a low-cost-to-serve customer segment, yet no viable business model has emerged around these types of users. Google and Facebook perfected advertising-based consumer models in the previous era, but this new market hasn't figured out monetization. Advertising remains the fundamental consumer business model because consumers resist paying directly for software.
Two truths Silicon Valley must relearn every decade: consumers do not want to pay for software, and consumers do not care about being productive. The canonical example is Dropbox, which built an exceptional product that grew rapidly but then faced a two-year lull. Drew Houston wanted to build a consumer company, but not enough consumers would pay for the service. The company had to completely rebuild its app from the ground up to serve enterprise needs—permissions, control, administrative features—because enterprises pay for productivity gains while consumers want to relax after work rather than be more productive.
OpenAI has repeated Dropbox's story at 100x scale by attempting to sell subscriptions to consumers. While they sold subscriptions, it wasn't enough volume. The company is now pivoting to advertising, rolling out ad features rapidly including copy capabilities and retailer connections for purchase tracking. Had OpenAI leaned into advertising immediately when ChatGPT became a hit, they would likely have a stronger ad product today, putting Google and Meta in more significant competitive trouble.
The key advantage of advertising-based monetization is that price increases get passed to advertisers rather than consumers, eliminating elasticity issues. Subscription models like Netflix face consumer rebellion when raising prices, making it difficult to increase revenue without losing customers.
Current estimates project $800 billion in capex spending this year, rising to $1.3 trillion next year. All installed compute capacity gets utilized immediately upon coming online, creating unusual market dynamics. The compute shortage stems from insufficient investment in 2023 and 2024, and TSMC's decreased growth rate during this period means the shortage will worsen through 2028-2029 because fab construction lead times exceed data center development timelines.
Andy Jassy and Satya Nadella describe building data center shells without immediate GPU purchases until demand materializes. However, once the fixed cost investment is made in physical infrastructure, companies won't let that capital sit idle. This mirrors commodity market logic where tech companies traditionally struggle to understand dynamics—tech focuses on differentiated products with pricing power, but commodity markets price based on marginal supplier costs.
In commodity markets, the marginal cost to serve determines pricing. The shipping industry provides the clearest example: once a ship is purchased, depreciation represents a sunk cost, and marginal costs (fuel, crew, port fees) are relatively low. Ships will operate at whatever market rates cover marginal costs, driving container prices down to the lowest sustainable level. When prices fall below actual cash operating costs, capacity exits the market, eventually restoring pricing balance.
During COVID, container rates surged from typical $3,000-$4,000 levels to $17,000-$18,000 due to supply shortages, generating extraordinary profits for shipping companies. However, the two-year ship construction timeline means simultaneous investment decisions create future oversupply when new capacity arrives together.
Memory markets exhibit classic boom-bust dynamics where capacity additions lag demand signals. Current AI investment follows the same pattern—companies measure payback periods during scarcity conditions, but these calculations may not hold during abundance. The concern is that massive capital deployment now may not generate sufficient revenues before capital markets lose patience, even if AI's long-term economic impact proves transformative.
Memory manufacturing has consolidated from many players to an oligopoly of three major companies. Historical cycles saw new entrants during booms, followed by price collapses and company failures. Samsung's market dominance came from investing counter-cyclically during downturns—requiring significant capital, discipline, and willingness to accept short-term losses for long-term market position. This strategy effectively eliminated Japanese competitors.
The current three-player structure creates more disciplined capacity management, avoiding past boom-bust mistakes. However, Apple is lobbying for Chinese memory suppliers, and algorithmic improvements focus on reducing memory requirements, potentially undermining the memory makers' market position.
TSMC represents the extreme case of concentrated risk, being the sole leading-edge foundry. The company has systematically transferred capacity risk to big tech customers by maintaining conservative expansion policies. TSMC's 30-year fab lifecycle planning makes them extremely cautious about overbuilding, creating a culture that prioritizes avoiding excess capacity over meeting immediate demand.
Morris Chang exemplified the counter-cyclical approach when he returned from retirement during the Great Recession to reverse spending cuts and invest aggressively as the iPhone launched. This decision laid the foundation for TSMC's leading-edge semiconductor dominance.
Risk doesn't disappear when transferred—it manifests differently. TSMC's conservative approach means big tech companies forego revenue and profits from insufficient compute availability. The 5G investment wave around 2020-2022 left TSMC positioned for that market, but AI demand emerged differently, requiring renewed capacity expansion that won't fully materialize until several years after demand signals became clear.
The extreme compute shortage creates conditions where big tech companies will endure the pain of developing alternative foundry relationships with Intel and Samsung. The scarcity itself provides the economic incentive for this transition. Intel's fundamental challenge has been cultural and organizational—they lacked customer service mindset, IP building blocks, and the service orientation that makes TSMC successful. However, current market conditions may finally provide sufficient motivation for major tech companies to invest in bringing Intel's foundry business to viability.
Amazon consistently builds products first for their own massive internal needs, then externalizes successful capabilities. AWS wasn't built on spare capacity—Amazon.com migrated to AWS later. The retail operation's scale necessitated building scalable, API-driven compute infrastructure that required no meetings or human interaction, which then became the foundation for serving external customers.
Similar patterns appear in logistics, where Amazon initially relied on external providers before building internal capabilities now offered to third parties. Early versions of Amazon's Graviton and Trainium chips were inferior, but internal usage provided the scale and feedback necessary for improvement.
Amazon uses managed services like Redshift where customers don't see the underlying processor. This allows Amazon to run their own custom chips underneath services, giving them volume and capacity to iterate and improve the chips before selling them externally. Graviton became better because Amazon was the first and best customer, and the same happened with Tranium, which now runs Anthropic. Amazon is applying this same approach to AI products.
Amazon has call center software and customer experience tools running through AI. The chatbot experience has improved significantly in the last year, where users are clearly talking to a chatbot but it effectively solves problems. Amazon is building AI services for their own business that they'll make broadly available, with some succeeding and others failing.
Amazon's core business feels impervious to AI disruption while benefiting from AI capabilities. Their investments in the real world create a deeper moat than competitors, and their ability to generate new business lines organically is compelling.
Apple has sat out the AI boom and appears to benefit from being lucky rather than good. Apple owns access to customers through their ecosystem, enabling the classic aggregator play where suppliers come to them rather than the other way around. Apple can acquire AI suppliers as needed.
Apple can serve users a chatbot experience, and with improved Siri functionality, there's potential for on-device AI that doesn't require paying inference costs since customers provide the electricity. However, Apple currently uses Google Cloud and Nvidia chips, suggesting they're not yet at full on-device capability.
Apple's physical goods business provides insulation - making phones is difficult, and having retail distribution for physical goods is valuable. The smartphone remains the perfect form factor: small enough to fit in a pocket, large enough to watch content, capable of running entire lives and entertainment.
The question is whether the phone will always remain the center of computing, or whether ambient AI in the home represents a different paradigm. OpenAI's efforts in this space are interesting for creating ambient AI where users can simply talk to AI and receive what they need. Apple is best positioned to provide ambient AI, but questions remain about whether they can do so without leading-edge models or if their phone-centric approach creates limitations.
Microsoft's mobile failure provides a cautionary tale - they were early to mobile but assumed the PC would always be central, creating phones that were essentially small PCs. Apple realized the phone needed to be the center, not an accessory to the Mac. The question is whether Apple falls into a similar trap by assuming the phone remains central, or whether ambient AI and cloud-based AI everywhere becomes disruptive to their position.
AI is a probabilistic endeavor, while Apple excels at deterministic products. When Apple ships an iPhone, it must work perfectly because failures cost billions of dollars. Apple has never had an iPhone recall, which demonstrates their careful decision-making, supply chain management, and hard decision-making processes. These qualities differ significantly from what makes great AI. Companies should focus on what they're good at, so Apple continuing to make great devices rather than pursuing AI makes sense.
The five potential frontier AI winners are OpenAI, Anthropic, Gemini, xAI (Grok), and Meta. OpenAI and Anthropic are the riskiest but have the biggest upside. The power of belief shouldn't be discounted - the most impactful historical events have been fueled by religion, and the two religious organizations in Silicon Valley are OpenAI (mainline evangelicals who go to church every Sunday) and Anthropic (all-in believers). This belief goes a long way, as does the need to make a business work for survival.
Google only needs search to not die too quickly. Meta has a huge advertising business, and in a world where Meta was run by anyone other than Mark Zuckerberg, they wouldn't be on the leading edge. This represents one of the purest manifestations of founder energy. Meta's business is so strong they can easily double down on AI. Google has been doing research in AI, making their pursuit logical, but Meta hiring a completely new team and starting from scratch is remarkable.
xAI's data centers in space represent a highly differentiated play. If this strategy succeeds, they may not need to own their own model to execute it, raising questions about why they're spending billions on model development in the meantime. The cursor acquisition makes tactical sense for both companies.
Meta is probably the most interesting case because there's a strong argument that it's reckless for digital companies not to be on the frontier. The counterexample is Microsoft, which isn't on the frontier but generated $20 billion in free cash flow last quarter and paid a $10 billion dividend. Microsoft's strategy is to play all models off each other, provide middleware and enterprise platforms, and disintermediate the models.
Microsoft's approach echoes IBM's strategy from the 1990s. IBM was a dominant company in the 1970s but became distressed by the 1990s. Lou Gerstner's key insight was that IBM was mediocre at everything - the price of monopoly is losing competitive capacity because success removes the pressure to excel. Once that muscle atrophies, it's gone.
Gerstner realized breaking up IBM would be disastrous because the component pieces weren't good individually. IBM's asset was being big at a time when companies needed help getting online. IBM built a huge consulting force and middleware layer between old mainframes and modern web services, helping corporate America get online. This gave IBM a 30-year lease on life.
Microsoft's playbook is similar - they'll help companies figure out AI without giving away crown jewels, building a platform and middle layer that manages changing models while providing backwards compatibility. Middleware saws off sharp edges, providing lowest common denominator capacity, but offers dependability and stability.
Microsoft's strategy is sound but also desperate in an existential way. The question is why companies use Microsoft products - because they did before. AI could potentially replace the tedious, repetitive work of moving systems of record. While Microsoft has some systems of record like Dynamics, their core value is user interface and interaction with computers. Tools like Codex and Claude Co-work target exactly what Microsoft has.
In the long run, all digital companies are threatened. Meta is a digital company with software, though not immediately threatened. AI takes up more time, and time is Meta's currency. OpenAI's Sora didn't take off, and social networks are hard and expensive to build.
YouTube pays creators from the beginning, creating marginal content costs. Unlike Netflix, which pays upfront for content, YouTube shares revenue after the fact. Meta pays nothing for content - Instagram generates enormous value while costing Meta zero dollars for content.
AI-generated content could theoretically benefit YouTube if inference costs are less than creator payments. For Meta, AI-generated content would have worse margins than their current free content model. However, in a world of constant AI interaction, the desire for human connection increases, potentially returning value to social networks.
Meta made a mistake with Snapchat by layering Snapchat's features onto their network rather than competing directly. TikTok represents a blind spot because it's classified as a social network but is actually an entertainment product. TikTok doesn't matter who you follow - content is determined by what you've watched, pulling the best content from across the entire network rather than limiting to social connections.
The biggest monetization impact from models is probably the incremental gains happening for Google and Meta, not OpenAI or Anthropic. Meta and Google can validate their image and text creation through their ad marketplace by running millions of A/B tests. Most ads don't convert, so they have a massive advantage with a liquid market that serves as a verification machine where humans decide whether to click and purchase.
This creates a feedback loop making their products better. Ad matching currently uses crude vector calculations, but language models could predict what people want to see next and find relevant content. A few percentage points improvement in ad relevance could generate billions of dollars. This alone justifies Meta investing in leading-edge models.
Meta's problem is they don't tell this story effectively. Mark Zuckerberg shares Sam Altman's problem of not loving ads, despite having the best ad business in the world. Meta's advertising is a societal positive - connecting products with people who didn't know they wanted them, creating happy customers and new business for entrepreneurs, all while providing free entertainment.
Mark Zuckerberg has never really discussed advertising's societal benefits except in passing. When Sheryl Sandberg left, the consistent communication about advertising's value stopped. Apple passing ATT (App Tracking Transparency) was one of the worst antitrust violations in technology history - Apple unilaterally destroyed business models while building their own advertising tracking. Apple's bus advertisement about companies overhearing conversations was dishonest about how internet advertising works.
Meta wasn't prepared to respond because they lack someone consistently making the case for advertising. The company seems embarrassed about their ad business, emphasizing glasses and AI instead. If Meta had communicated advertising's benefits more consistently, they would be in a better PR position, have better standing relative to Apple, and find it easier to convince Wall Street to allow investment.
Meta spent cumulatively over $100 billion on Oculus, which was dated from the beginning, creating questions about why they should be allowed to spend money again.
Jensen Huang and Nvidia represent the major player not yet discussed extensively. The question is whether compute will ultimately be a commodity and whether intelligence will be a commodity. The internet demonstrates how commodities change the world through free distribution on a marginal cost basis. Bandwidth is a commodity, yet the internet changed the world because it's available to everyone. Differentiated products by definition have lower TAMs because of price elasticity constraints. Apple's device sales will never serve the whole world the way Google can through free services. Commodities are impactful to the extent they're available to everyone.
Nvidia's position appears unnatural given they've maintained margins despite everyone coming for them in 2026. However, Nvidia isn't actually maintaining margins when accounting for circular financing arrangements. Nvidia provides a 25% backstop and takes equity in Neo clouds, guaranteeing they'll buy compute through 2030. This allows entities to get lower cost of capital to buy more GPUs. Nvidia assumes risk that has a price, and this risk-taking represents an expected value calculation that's neither zero nor 100%, effectively functioning as a price cut that doesn't appear in traditional margins.
Nvidia's ultimate competitors are the hyperscalers, particularly Google and Amazon, who are building their own chips and selling them externally. Google made a deal to sell 20% of their TPUs to Anthropic. Amazon is preparing to sell Trainium 3Es or potentially Trainium 4 externally. The hyperscalers sell chips as commodities, not on differentiation, which means they're not cannibalizing their cloud attractiveness. This represents Nvidia's biggest problem because the hyperscalers have lower cost of capital than Neo clouds, creating a capital fight.
CUDA's moat has diminished because models don't care what they run on, though it's still something of a mode. Nvidia pursues sovereign clouds and enterprises because these markets don't balance chips versus competitors the way hyperscalers do. Intel was cleaned out of hyperscaler sales by AMD because hyperscalers invested in making AMD work, but Intel maintained government and enterprise sales where customers lack resources to switch. Nvidia's sovereign cloud strategy targets markets where customers won't compare chips.
Nvidia may benefit if power becomes constrained, as everyone would want the most token-efficient chips, where Nvidia maintains an advantage. The US has brought more power online than expected through behind-the-meter solutions replicated in West Texas with natural gas, plus nuclear plant restarts. The longer power remains abundant, the more time Amazon has to improve Trainium and Google has to make TPUs competitive on efficiency, making Nvidia's margins harder to sustain.
The .com bubble produced lasting benefits including fiber in the ground that Google purchased as dark fiber after the era. The core internet still runs on WorldCom fiber. BNSF railroad continues generating revenue for Google from Northern Pacific and Jay Cooke selling bonds to retail investors. The question becomes what lasting benefit AI will produce, with power abundance being the most valuable potential outcome given historical energy constraints.
We've always been energy constrained, making it difficult to imagine energy abundance. Power constraints are real but have taken longer to manifest than expected, surprising even Jensen Huang who anticipated power insufficiency as Nvidia's moat sooner than it occurred.
Keep Invest Like The Best in your library
Save the videos and channels worth coming back to, and find them again in one place.
